U.S. patent number 3,993,976 [Application Number 05/469,194] was granted by the patent office on 1976-11-23 for method and apparatus for pattern analysis.
This patent grant is currently assigned to The United States of America as represented by the Secretary of the Air. Invention is credited to Arthur P. Ginsburg.
United States Patent |
3,993,976 |
Ginsburg |
November 23, 1976 |
Method and apparatus for pattern analysis
Abstract
A technique and apparatus for two dimensional pattern analysis
utilizing a transform of the pattern enables the extraction of
desired pattern information by means of spatial filtering in
accordance with known human visual system processing. Two
dimensional spatial frequencies resulting from the transform are
acted on by either anisotropic or uniquely used conventional
filters to extract one, two and three dimensional pattern
information from spatial frequency subsets to determine general
form, edge, texture and depth information for detection,
identification and classification of objects in simple or complex
scenes.
Inventors: |
Ginsburg; Arthur P. (N.
Billerica, MA) |
Assignee: |
The United States of America as
represented by the Secretary of the Air (Washington,
DC)
|
Family
ID: |
23862824 |
Appl.
No.: |
05/469,194 |
Filed: |
May 13, 1974 |
Current U.S.
Class: |
382/211; 359/559;
382/280; 250/550 |
Current CPC
Class: |
G06K
9/748 (20130101) |
Current International
Class: |
G06K
9/74 (20060101); G06K 009/00 () |
Field of
Search: |
;340/146.3P,146.3E
;350/162SF,3.5 ;250/550 |
References Cited
[Referenced By]
U.S. Patent Documents
Other References
Stromeyer III et al., "Spatial-Frequency Masking in Vision,"
Journal of the Optical Society of America, (vol. 62, No. 10), Oct.
1972, pp. 1221-1232. .
Schneider et al., "Spatial Frequency Range Scanning Using a Zoom
Objective," Applied Optics, (vol. 11, No. 8), Aug., 1972, p.
1875..
|
Primary Examiner: Boudreau; Leo H.
Attorney, Agent or Firm: Rusz; Joseph E. Goldman; Sherman
H.
Claims
I claim:
1. Apparatus for analyzing a two dimensional pattern
comprising,
means for transforming the pattern into a spatial frequency domain
to relate harmonically the elements forming the pattern
information,
means for filtering the transformed pattern substantially in
accordance with the anisotropic attenuation characteristics of the
human visual system, and
means for bandpass filtering the filtered transform to extract
spatial frequency subsets from the said filtered transform for
feature analysis.
2. An apparatus as defined in claim 1 wherein said feature analysis
is obtained by correlation of said spatial filter subsets with
stored information or for inverse transform and reconstructing band
spatial frequency subsets for further analysis.
3. A static, single stage, two dimensional, anisotropic, spatial
frequency magnitude filter having spatial frequency attenuation
characteristics corresponding substantially to the average contrast
sensitivity of the human physiological visual system resolution at
particular ambient light conditions over 360.degree. of viewing
angle.
Description
BACKGROUND OF THE INVENTION
The invention relates generally to pattern analysis and more
particularly to two dimensional pattern analysis performed by
attenuating and isolating spatial frequency subsets of a transform
of the pattern. Transform attenuation corresponding to the
variations of human contrast sensitivity over 360.degree.
orientation may be used to bias two dimensional transform data such
that one, two and three dimensional pattern information in spatial
frequency subsets may be extracted by using conventional bandpass
spatial filters. In addition, the conventional bandpass spatial
filter may provide depth information by uniquely using it in
combination with the transform with or without the transform
attenuation corresponding to variations of human contrast
sensitivity.
The prior art has attempted similar two dimensional pattern
analysis using contrast sensitivity data; however, any contrast
sensitivity attenuation attempted was performed isotropically over
360.degree.. Also, special binary valued, wedge-type filters were
used to obtain spatial frequency information at various
orientations. However, these devices required complex means for
extracting the pattern elements to create special frequency
signature rather than enabling the ulitization of simple
thresholding to obtain the same result for many tasks. Basic
pattern information previously required heuristic low pass bandpass
spatial filters. None of the prior art devices were able to obtain
large amounts of relevant pattern information from two dimensional
spatial frequency information because the attenuation
characteristics for orientation and spatial frequency and relevant
spatial frequency bands were not recognized or determined.
Furthermore, third dimensional depth information in terms of
spatial frequency subsets and reconstructed intensity gradients has
been beyond the realm of the two dimensional pattern analysis
systems heretofore contemplated. Finally, two dimensional pattern
analysis previously developed does not have the unification or
parsimony of the methods outlined herein.
SUMMARY OF THE INVENTION
A method and apparatus for two dimensional pattern analysis
utilizing two spatial filter types is presented. The first spatial
filter type is unique and attenuates the two dimensional pattern
transform data corresponding to human contrast sensitivity values.
This represents the human physiological visual system asymmetric
resolution of pattern information over 360.degree. of orientation.
The asymmetric spatial filter provides a great amount of pattern
information since it is orientation sensitive, especially in
automatically separating similar texture elements and forms from
backgrounds that differ only in orientation. The attenuation
characteristics of the asymmetric spatial filter automatically
provides, upon reconstruction of the attenuated transform,
intensity values that allow the use of simple thresholding to
isolate similar pattern elements differing in orientation over a
continuum of intensity values. The second spatial filter type is a
conventional bandpass filter. The attenuation characteristics of
the anisotropic spatial filter allows basic pattern form
information to be extracted by a conventional bandpass spatial
filters based upon energy of the transform components resulting in
less transform information to be stored in a memory and processed
for the classification scheme. Translating conventional spatial
filters over the two dimensional transform to isolate spatial
frequency subsets, according to energy contained therein, renders
possible the extraction of three dimensional information by
correlating that information with similar information stored in
memory or by retransforming just the isolated spatial frequency
subset. Third dimension depth information is in terms of pattern
intensity with concomitant shape changes can be seen explicitly in
a retransformed pattern. This technique of translated spatial
filters also allows the extraction and biasing of selected original
pattern edge features in the reconstructed pattern. In addition,
the very low spatial frequency information, e.g., the fundamental
spatial frequency also is used to extract third dimensional depth
information over large pattern areas, e.g., depth information from
different texture gradients.
Thus, it is the primary object of this invention to develop a
method and apparatus that enables the isolation and extraction of
edge, form and texture, and three dimensional (depth) information
from two dimensional patterns within the context of human visual
information processing.
It is an object of this invention to obtain edge information from a
two dimensional display by the translation of a conventional
bandpass filter over two dimensional spatial frequency subsets.
It is another of this invention to obtain shape changes and
intensity gradients from a two dimensional display which are
correlated to patterns in depth.
It is still another object of this invention to provide a system
utilizing a two dimensional Fourier or other transform which is
coupled to a spatial filter having a modulation transfer function
corresponding to human contrast sensitivity values and a
conventional spatial filter to get improved correlation with the
human visual system processing for more accurate selection of
important spatial frequencies to be extracted by the conventional
bandpass spatial filters.
It is a further object of this invention to obtain edge, form,
texture and depth information from a two dimensional transform
which is filtered, wherein the transform and filter may be either
digitally or optically generated.
It is a still further object of this invention to provide a method
and apparatus for two dimensional pattern analysis which
corresponds to that performed by the human visual system in that
spatial filtering is made to correspond to the human physiological
visual system contrast sensitivity.
Another object of this invention is to obtain a two dimensional
pattern analysis which utilizes a transform of the pattern into
bands from which subsets may be extracted.
Still another object of this invention involves the obtaining of
form information from a two dimensional pattern with a minimal
amount of spatial frequencies.
A further object of this invention involves two dimensional pattern
analysis wherein a transform is obtained having the fundamental
frequency utilized for intensity information relating to depth; the
mid-low frequencies providing form and edge information; while the
high frequencies of the transform provides fine details.
A still further object of this invention involves the provision of
a modulation transfer function (MTF) spatial filter for use with a
transform of a two dimensional pattern wherein the MTF filter
biases the frequencies of the transform in accordance with human
contrast sensitivity with anisotropic characteristics.
It is another object of this invention to provide a two dimensional
pattern analysis system whereby orientation sensitivity of a
spatial filter provides edge information and specific form
extraction.
It is still another object of this invention to provide two
dimensional pattern transform analysis techniques. Although some
similar analysis can be performed mathematically in the space
domain, space domain techniques such as convolution lose space
invariance (translational invariance) and analytical power for
separating information.
It is a further object of this invention to provide a technique and
apparatus capable of automatically detecting classifying and
identifying targets contained in a two dimensional data processors
such as is provided by high resolution radar imagery.
It is a still further object of this invention to provide an
apparatus for pattern analysis which is easy and economical to
produce comprised of conventional currently available materials
that lend themselves to standard mass production manufacturing
techniques.
These and other advantages, features and objects of the invention
will become more apparent from the following description taken in
connection with the illustrative embodiment in the accompanying
drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a block diagram illustrating the various steps and
apparatus required for performance of two dimensional pattern
analysis in order to obtain edge, texture, form and depth;
FIG. 2 is a two dimensional intensity pattern of a letter G formed
by dots;
FIG. 3 is the upper half of a Fourier transform matrix representing
the amplitude of the spatial frequency components in the two
dimensional transform plane of the dotted letter G of FIG. 2;
FIG. 4 is the normalized Fourier transform of FIG. 3;
FIG. 5 is a typical modulation transform function set of values for
a 64 by 64 array;
FIG. 6 is a digital representation of the normalized amplitude
spectrum of the dotted G represented in the Fourier transform of
FIG. 4 after it has been acted on by the MTF filter of FIG. 5;
FIG. 7 is the inverse transform of FIG. 6 (with the bottom half
presented) in order to illustrate the reconstructed dotted G and
the effects of the MTF filter on the original dotted G;
FIG. 8 is a digital representation of the dotted G representation
of FIG. 7 after further spatial filtering by conventional low pass
spatial filter; and
FIG. 9 is a block diagram illustrating target detection and
identification techniques.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
The technique and apparatus for two dimensional pattern analysis is
illustrated below in conjunction with specific Figures to which
reference will be made. It is contemplated that both feature
extraction, detection and classification will be encompassed within
the pattern recognition aspects of the invention. Since feature
extraction techniques of the human visual system have been
investigated, the technique and apparatus of this invention is made
to correspond to the human visual system. The correspondence of the
apparatus of this invention to the human visual system has been
provided by allocating the processing devices of the human to
specific apparatus. For example, the visual system is divided into
three basic units, i.e., preprocessing, transforming and
classification. The preprocessing concerns the processing of the
pattern, p(x,y), from the lens and retina through the primary
visual cortex, which would be represented in the apparatus of this
system by a two dimensional, anisotropic, spatial filter (MTF) -
H.sub.MTF (u,v) -. The transforming step would be performed by a
two dimensional Fourier transform P(u,v) which operates on the
preprocessed data. Classification is obtained by the selection of
subsets of the Fourier transform by means of conventional bandpass
spatial filter H.sub.BP (u,v) techniques that are correlated to
stored spatial frequency subsets of patterns for detection
classification or identification. The reconstructed pattern,
P.sub.r (x,y), is used for further analysis and thresholding
techniques.
In the foregoing paragraph p(x,y) is the two dimensional intensity
distribution of the pattern where x,y are the space coordinates in
the two dimensional pattern plane. H.sub.MTF (u,v) is the percent
contrast sensitivity values over the two dimensional transform
plane (u,v). ##EQU1## H.sub.BP (u,v) is a representation of the
bandpass filter where selected values of u,v would be equal to one
for passed or correlated spatial frequencies and all others would
be equal to zero. The pattern ##EQU2## where X = 2u + 1; Y = 2v + 1
and ##EQU3## are the spatial frequencies in cycles per unit length,
comprises the reconstructed two dimensional pattern after the
Fourier transform and both MTF and bandpass spatial filtering.
Utilization of the principles of this invention can be achieved
with respect to an understnading of the block diagram of FIG. 1. In
this Figure there is shown a two dimensional pattern at 10 which
may be either digitally or optically presented, for example, as two
dimensional radar data or a photograph. The transform at 12 results
in separate frequency bands collected into frequency subsets, each
of which contains specific pattern form or varieties of pattern
form information. The transform may be performed by conventional
optical means or by means of electronic circuitry such as a digital
computer. The discussion of this invention will be related to a
centro-symmetric Fourier transform since this is the easiest type
to visualize and is found in the transform plane of an optical
system. Other types of transform may be used and, like the two
dimensional pattern, may be either optically or electronically,
i.e., digitally represented. The transform 12 takes the two
dimensional intensity information from the pattern at 10 and
analyzes it to a multiplicity of frequency bands. The high
frequency bands contains the fine detail while the low frequency
bands contain gross form. When digital Fourier transforms are used,
a numerical printout describes discrete frequencies whereas an
optical Fourier transform provides infinite spatial
frequencies.
Spatial frequency is a single value comprised of a real and an
imaginary part whose magnitude ##EQU4## is represented as a
function of position is the u,v transform plane. The central
spatial frequency D.C. represents the average energy across the
pattern.
Each band increases in spatial frequency as one goes from the
center of the transform (D.C.) and the same spatial frequency value
is in a radially concentric pattern in an optical system or in a
square shape pattern in a digital rectangular coordinate system.
Collections of either one or more of the spatial frequencies in the
frequency bands equally or unequally over the two dimensional
transform plane are called subsets.
The MTF (Modulation Transfer Function) spatial filter is shown at
14 on FIG. 1 and is constructed so as to bias or attenuate
particular frequencies at particular orientations. The attenuation
is asymmetric and corresponds to the average value of (percent)
contrast sensitivity for various angles in accordance with the
human physiological visual system resolution. The MTF filter data
ulitized in the pattern analysis of this invention relates to
0.degree., 45.degree. and 90.degree. angles with interpolation
providing for values for orientations other than those obtained by
measurement.
The spatial filter at 16 would be a conventional bandpass filter
either isotropic or anisotropic depending upon its use as will be
explained infra.
Examples of the items described relative to the diagram of FIG. 1
will be illustrated with respect to FIGS. 2-8. The two dimensional
intensity pattern of the letter G formed by dots is digitally
represented in FIG. 2. The blanks represent zero intensity, and
digits from 1 through 9 are given their intensity value. In this
and other digital representations, the value of 10 would be printed
out as zero and values 0.1 to 1 would be represented as a dot.
The transform 12, the lower half of which is omitted to avoid
duplication since the transform is symmetric, is illustrated in
FIG. 3 and provides a digital representation of discrete Fourier
amplitude transform values of the pattern of FIG. 2. FIG. 4 is a
digital representation of the normalized transform of FIG. 3 also
with the lower half removed for simplicity. The MTF spatial filter
14 of FIG. 1, which is always anisotropic and is represented
digitally in FIG. 5 for only one quadrant of a 64 by 64 array,
since the remaining quadrants can be achieved by appropriate
replication in the other three quadrants. It should be noted at
this point that the MTF filters can have different values depending
upon the data obtained from the human visual system. That which is
shown is for high sensitivity at low spatial frequencies. When the
two dimensional pattern of FIG. 2 has been transformed and
normalized as illustrated in FIGS. 3 and 4 and had the MTF filter
of FIG. 5 applied thereto, the resultant of FIG. 6 provides a
representation of the attenuated, normalized amplitude spectrum of
the dotted G. The reconstructed G pattern by the inverse Fourier
transform is illustrated in FIG. 7 which demonstrates the
application of the MTF filter on the original G pattern. Note that
the dots are still resolved and that the blurring corresponds to
what an observer reports when viewing a photograph constructed from
that data. Further conventional bandpass spatial filtering which
allows only the first four spatial frequencies to be used in
reconstruction of the G pattern by a Fourier transform results in
FIG. 8 where a general G form inherent in the dotted G form has
been extracted.
The dot letter G for which examples were given in FIGS. 2 through 8
was related to the obtaining of general G form information from the
two dimensional transform. Edge, texture and depth information may
also be obtained from transform data. In order to obtain edge
information we would use the pattern 10 of FIG. 1 to which the
transform 12 has been obtained, and energy normalized (divide each
spatial frequency by the D.C. term). At this point the MTF spatial
filter 14 could be utilized to bias the edges according to
orientation; however, the output from the transform could be sent
directly to a conventional mid to high frequency bandpass spatial
filter 16 for unbiased edges according to orientation. This would
result in a pattern outline upon reconstruction. When the MTF
filter is utilized, its output could go to the conventional spatial
filter 16, which is a mid to high frequency bandpass filter, to
extract spatial frequencies for correlation with stored spatial
frequencies in a memory or classification device 18. This would
allow detection and identification based upon edge pattern
features. If the mid to high frequency bandpass filter is isotropic
it would best provide form edge outline, whereas anisotropic
filters are used for edges at desired orientations. Translation of
the spatial filters over the transform plane according to energy in
frequency orientation subsets may be used in order to obtain edges
at one, two or more orientations. When the mid to high frequency
bandpass filter is used for reconstruction, the inverse transform
is used instead of the correlation as previously defined. The
transform, MTF and bandpass filter steps would be the same as
previously described, however, the inverse transform as illustrated
at 20 of FIG. 1 would provide highlighted edges and would allow the
use of a threshold device 22 which would select intensity ranges by
values to separate highlighted edge information from other pattern
information. The threshold edges may be used for inputs for form or
texture processing. If the transform MTF filter is reconstructed by
an inverse transform determination and effects and pattern
resolution of the human visual system physiological filtering are
obtained.
To obtain form we would, as described by the edge information
processing, transform the pattern of 10 at 12 and energy normalize
it and also use an MTF filter to attenuate the background from the
form or skip the MTF step for unattenuated background. Here we
would use a low frequency bandpass filter 16 without the
fundamental or a few of the lowest frequencies to obtain the basic
form information for correlation at 18 with similar stored spatial
frequencies in the memory to detect and identify the form. If the
filter 16 is isotropic and translated according to the energy in
the frequency orientation subsets, three dimensional forms in
memory may be correlated to forms in depth. Reconstruction would
utilize the low frequency bandpass filter where the image is
transformed, spatially filtered by the MTF filter and applied to an
inverse transform 20 to reconstruct patterns where the basis for
correlation here would be observed from human analysis or threshold
techniques that provide a separation of form information from other
pattern information to be provided at 22. To determine the effects
and pattern resolution of a human visual system physiological
filtering one would transform the image after MTF spatial filtering
and inverse transform.
Texture analysis is divided into two systems. Where the texture
elements diffier in slope, intensity and in size one would apply a
transform 12 to the two dimensional pattern 10 and normalize it,
apply the MTF filter to form texture clusters and elminate the MTF
filter where there is no texture cluster formation desired. A mid
high frequency bandpass filter 16 is used to extract spatial
frequencies for correlation with the classification system 18. An
isotropic filter would be used for texture band whereas an
anisotropic spatial filter would be used for texture clusters
differing in slope. Reconstruction by an inverse transform at 20
could be achieved when a transform is applied to an image MTF
filtered. The inverse transform is used to reconstruct the filter
pattern so that texture clusters are formed for human analysis or
threshold techniques could be used for separating texture clusters
from other pattern information. The cluster forms may be used as
inputs for form processing.
When texture elements differ only in slope or shape from other
forms here we would transform an applied MTF filter to attenuate
texture and form elements that differ in slope and shape or
eliminate this step where no pattern segregation exists in terms of
slope. The conventional spatial filtering of the mid high frequency
range here would extract spatial frequencies for correlation with
stored spatial frequencies in a memory for detection and
identification. An isotropic spatial filter would be used after the
MTF filter whereas an anisotropic spatial filter would be used to
extract texture element according to the energy at various
orientations. The mid high frequency bandpass filter could have its
output applied to the inverse transform for reconstruction when the
MTF has been used. The inverse transform allows for observation of
texture elements differing in intensity according to slope and
shape. However, threshold techniques which isolate texture elements
according to slope and shape by intensity variations could be used
to segregate these texture elements and provides an input for
further processing. Determination of effects of pattern resolution
of human visual physiological filtering could be obtained by
utilizing the transform 12 spatial filter 14 and inverse transform
20.
Depth information in planes over large window areas uses the
transform 12 on the pattern 10 and may optionally use the spatial
filter 14 where a low frequency bandpass filter at 16 removes the
fundamental frequency for correlation at 18 with stored spatial
frequencies in a memory to determine if depth planes are present.
Anisotropic filters would be used for oriented depth planes whereas
isotropic bandpass filters would provide the usual depth
information. If construction via an inverse transform 20 is desired
the transform at 12 with or without step 14 would be applied to the
inverse transform. The phase terms of the fundamental spatial
frequency or higher frequencies are used to phase lock high spatial
frequency information over selected depth planes for form
isolation. Here we would select high frequencies having the same
phase terms as determined by the very low spatial frequencies.
Thus, the inverse transform would reconstruct the filter pattern
and the depth plane and forms could be observed for human analysis
or by utilizing the threshold techniques illustrated in the flow
diagram in block 22.
Where the depth information is a function of pattern intensity
variations of smaller forms we would apply the transform 12 to the
pattern 10 to obtain an energy normalized pattern with or without
the MTF filter. A mid to high frequency bandpass filter would be
used, isotropic or anisotropic, translated over high energy subsets
according to energy contained in the transform domain. Either
correlation of the spatial frequency with similar information
stored in memory or reconstruction via an inverse transform may be
accomplished for detection, identification, or threshold
techniques. The spatial filter bandwidths may be generally
determined from the half-power of the pattern elements, to be
detected, classified or identified, for example, a large form
filling a 32 by 32 pattern element window may use a bandpass filter
of f=2, 4 cycles per picture width to capture its basic form.
The detection, classification and identification of objects is
performed in the transform domain by comparing spatial frequency
subsets from the input pattern with similar information stored in
memory. This comparison may be done by using a matched spatial
filter or cross-correlating the input spatial frequencies with
stored spatial frequencies.
Matched spatial filtering, used primarily for detecting and
locating objects in complex scenes, is usually used for optical
processing whereas cross-correlation is usually used in the digital
processing. Those terms may be used for either optical or digital
processing and are equivalent, except that a matched spatial
filter, if followed by an inverse transform, results in an
intensity value whose value and position in the reconstructed plane
represent the degree of comparison and location of the detected
object.
Maximizing cross-correlation is equivalent to minimizing Euclidean
distance (d) between two spatial frequency subsets. ##EQU5## where
M & N are the number of spatial frequencies in the x and y
directions, respectively, used for comparison.
Re.sub.p = real part of transform of prototype in memory
Re.sub.i = real part of transform of the input pattern
Im.sub.p = imaginary part of the transform of the prototype in
memory
Im.sub.i = imaginary part of the input pattern
This Euclidean distance metric may be normalized and its value used
to rank order and thus provide a quantitative measure of how
similar the input pattern is to the pattern in memory.
These are linear classification schemes. Non-linear classification
schemes based on probability distribution or other decision
criteria may be used.
The phase lock technique is a unique method for segregating
connected objects in simple or complex scenes using the fundamental
or low spatial frequencies to determine subcomponents of a whole
pattern. For example, the patterns (assume a series of circles
divided equally and unequally by a straight line), each contains
different fundamental spatial frequency information directly and
results in different intensity distribution if inversely
transformed from the fundamental spatial frequency terms. (The
fundamental spatial frequency term is defined as one whole cycle of
the pattern width.) It is desirable to separate whole patterns into
subsections, for example, each section of the divided circle. This
is accomplished by correlating the fundamental spatial frequency
information with the pattern subsections. If a subsection exists
from that information, then by selecting all the higher spatial
frequency terms containing similar phase as that of the subsection
desired, one can obtain those isolated pattern subsections for
further analysis.
A basic flow diagram of the information processing techniques
proposed to solve the complex scene analysis or target detection
and identification problem is presented in FIG. 9. The
preprocessing, detection, and identification techniques are
discussed in detail. It should be stressed that although the
imagery will be initially processed digitally, these techniques,
especially the detection stage, may be optimally performed using
optical computers.
The preprocessing operations prepare the radar imagery for
subsequent target detection and identification techniques. The
radar imagery data must be made compatible to the processing
systems. Pre-digitized data stored on magnetic tape can be inputted
directly. Film imagery will have to be digitized by a scanniny
system. Image gray scales will be either linear or log normalized
to maintain uniformity of multi-source imagery. The log transform
will tend to enhance film contrast lost during original film
processing. Further image processing may be desired such as
de-convolution of speculars. The preprocessing operations may be
accomplished on the entire image data before additional processing
is undertaken or performed on each scan window before input to the
detection process. Some preprocessing steps could be possibly
eliminated if the detection and identification techniques are
performed on the image signal in either digital or film form.
The detection techniques are primarily designed to reduce
identification processing over image areas that are not of
interest. The central concept of the detection technique is to use
Priori inputs: desired targets to be detected; approximate target
size from radar imaging parameters; and target false alarm levels
to set up spatial matched filters for use in a cross-correlator.
Thus the detector will be looking only for desired targets that
exceed a predetermined target detection error.
A matched spatial filter (MSF) is known to be an optimal filter for
separating signals from noise in a linear system. In terms of radar
imaging, the unwanted targets and background clutter, i.e.,
terrain, are the noise to be separated from the desired targets,
the signals. The MSF has a transfer function with complex
transmittance proportional to the complex conjugate of the Fourier
transform of the signal. The MSF converts the complex wavefront of
the Fourier transform of the signal into a plane wave which is
focused into a bright point in the output plane of the correlator
whose position is directly related to the signal position at the
input plane. The spatial frequencies from the noise and undesired
targets will be attenuated at low valued spatial frequency regions
of the MSF, thus reducing the energy of these noise sources
detected at the output plane. This concept had been demonstrated
usually by detecting highly formatted patterns, e.g., a letter or
words repeated on a page of text. However, these demonstrations are
really biased and MSF techniques perform poorly when one uses
unformatted imagery. The simple reason for the failure is that
there is too much noise versus signal energy in most real world
scenes. There are two ways that are proposed to greatly reduce that
problem. Firstly, the aperture of the input plane of the correlator
could be reduced (or the image enlarged) to make the size of the
target large compared to the size of the aperture. In other words,
scan smaller image areas rather than correlate the entire image at
one time. Secondly, previous research has demonstrated that basic
shape information lies in a relatively small band of spatial
frequencies. Thus a bandpass MSF based upon gross target shape will
be generated and spatial frequencies generated by clutter, which
are primarily high spatial frequency data, will not be allowed to
pass into the output plane and reduce the signal to noise ratio.
The size of the scan window will be made as large as possible as to
allow for maximum detection areas with lowest tolerable signal to
noise ratios. Target size differences will be compensated for by
appropriate scaling of the bandpass MSF by extrapolation or
interpolation from the radar imaging system scaling parameters.
Target rotation in the image space produces a concomitant rotation
in the Fourier transform domain. Previous research with rotated
simple patterns using Fourier transform techniques demonstrated
successful pattern classification under .+-.30.degree. rotation.
Therefore, a solution to this problem is to rotate the MSF 360
degrees in 12 increments of 30 degrees each. Another solution is to
create one inclusive "OR" MSF from rotated MSFs to obviate any
rotation requirements during detection processing. The bandpass MSF
will still detect the basic target shape even though some target
degradation will occur due to radar aspect angle changes unless the
degradation is such that an observer could not detect the target.
Most target degradations due to aspect angle involve relatively
higher spatial frequencies. These techniques enable quick,
efficient target detection using MSF techniques that have
previously failed.
The energy content of the correlation peaks provide a degree of
similarity between the detected target and the prototype (MSF) as
well as positional information that will be required for the target
identification task. Correlation peak energy exceeding a
preselected false alarm level will indicate a detected target and
will be reporated as such. No targets detected in a scan window
will cause the same window to shift and detection processing to be
initiated over the new scan window.
The correlation peaks in the output plane of the correlator provide
the location of possible targets and lead the identification
processing to only those image areas of interest. Thus the more
detailed and time consuming processing elements are concentrated
only at selected target areas, an important consideration for any
possibility of real-time target identification in light of present
digital transform processing speeds. The central concept for target
identification is the same as that of the detection process except
that more shape (feature) information is used in a more controlled
decision process guided by the detection results.
A tapered grid is centered over the detected target whose size,
approximately 32 .times. 32 elements, will be a function of the
target size. The fast Fourier transform computed over the grid is
energy normalized and spatial filtered. The target spectral
components comprise the target feature vectors. Additional features
that have been previously determined to be required for correct
target identification can be extracted from other target spectral
components. Furthermore, target context, e.g., background-texture,
can be extracted from the higher spectral components. These
additional features can be added syntactically (isolated and put
into context) to the feature space which will be inputted to the
classifier/identifier discriminant. The main discriminant used
quite successfully for many diverse pattern classification tasks
has been minimum Euclidean distance, equivalent to maximizing
correlation, of low spatial frequencies with stored prototypes.
Other discriminants may be used. Previous pattern recognition tasks
have required at most 7 .times. 7 or 49 low spatial frequency
values to be stored as prototypes, a quite small amount of data
when compared to other feature extractor techniques such as
template matching. The stored prototypes will be the result of
averaged training set targets. It is important to realize that the
radar imaging artifacts such as shadowing effects are included in
the training sets and will not present any difficulty as long as
the shadowing effects are similar for the same target. Since the
images are a function of the target geometry, only variations due
to radar look angle will present deviations in the image. Thus the
technique presented here does not require optical fidelity of the
image but shape (low spatial frequency) fidelity of the imaging
system. It is emphasized that it is the low spatial frequency
target shape information that is used for these detection and
identification techniques.
Each detected target will be processed by the preceding techniques
and rank ordered against previously selected prototype targets. The
closest prototype in a Euclidean distance sense to the target will
be identified and reported as such with the normalized discriminant
value providing a measure of the degree of similarity between the
prototype and the target. It may be desired to use decision trees
to eliminate the different classes of prototypes to be classified
with any given target. For example, a swept wing aircraft can be
identified as such by off axis Fourier spatial frequencies which in
turn could exclude non-swept wing aircraft from the classification
process. Each detected target in each scan window will be processed
as just outlined until the complete image is processed.
Although this invention has been described relative to particular
embodiments, it will be understood that the invention is capable of
a variety of alternative embodiments. Each of the steps or
techniques may be performed optically or electronically. For
example, the translation can be optically performed by physical
movement or mathematically performed by electronic or symbolic
means. All of the elements are well known in the art and are
standard with respect to methodology or technique except for the
MTF filter, which makes information easier to obtain and use by
providing improved sensitivity for thresholding, the utilization of
translation for obtaining depth information, and phase comparison
with magnitude for complex scene or pattern analysis. The novel
combination of method steps and means provides results not
heretofore obtained. I intend to be limited only by the spirit and
scope of the appended claims.
* * * * *